PYTHON · NUMPY

The meter that sleeps

A sensor samples 25% of the time. The workload repeats with the same period as the sampler. How far apart can two reported averages be for the identical trace?

Show that a part-time power sensor can report two different answers for one identical workload, and say which measurement it can still be trusted for.

01Challenge

You have moved the meter to four planes and learned that where it sits changes the number. Now hold the plane fixed and change only when the meter is looking.

Predict firstA sensor samples 25% of the time. The workload repeats with the same period as the sampler. How far apart can two reported averages be for the identical trace?

02Model

The four previous lessons asked where the meter sits. This one asks when it is awake. A built-in sensor is convenient because it is already there, and the convenience hides a sampling geometry: it reports an average over the moments it happened to look at, not over the run. A 2024 study of more than seventy NVIDIA GPUs across twelve architectural generations found that on A100 and H100 only about a quarter of the runtime is sampled at all. That is survivable when power is flat. It stops being survivable the moment the workload has a period of its own, because then the sampler and the workload can beat against each other, and the reported number depends on their relative phase rather than on the energy. The useful part is that this is geometry, not noise: it is predictable, it is correctable, and an external meter or a longer, phase-randomised window recovers the truth.

03Guided practice

Write the reported average first: it is the mean over only the steps where the sensor is awake. Then run it at two phases, aligned and anti-aligned. The trace is identical in both; only the shutter moved. Read the two numbers against the true mean before you read the error column.

04Feedback

PASS WHEN You made one identical workload report 400 W and 100 W against a true 250 W, by moving nothing but the sampling phase.

If your run fails, check:

A sensor that samples part of the time reports the moments it saw, not the run.
Aliasing is geometry, not noise: it does not average away with a longer run at the same phase.

05Retrieve & space

Why does a longer run at the same sampling phase not fix this?
Your workload is flat. Does the 25% duty cycle still threaten your figure, and why?
You have only the built-in sensor. Name one change to the measurement that recovers the true mean.

06Mastery & project

You can state when a built-in power sensor is trustworthy and when it is not, in terms of the workload's period rather than the sensor's accuracy.

Take an energy figure you or someone else has published from a built-in sensor. State the workload's period, the sensor's sampling duty cycle, and whether the two are independent. If you cannot establish that they are, say so beside the number.

GO FURTHER · THE LIVE PLAYGROUND

PAI-350 · Reading an Energy Claim ↗ · The same question asked of someone else's published figure.
Open the instrument ↗ · Where the meter sits, which is the other half of this.

← The reporting record The counter that cannot miss →

PYTHON · NUMPY

The meter that sleeps

One duty-cycled workload, one part-time sensor, two phases. The sampling geometry is the whole model.

You read

the arrays and values already in scope

You change

the code you write in each cell

Fixed

the dataset and the checks that grade you